Loops HouseAgent shield

NEAR HACKATHON
Links
Team
1 member- MAOwner
Magnus
Overview
AgentShield lets an AI agent pay for things without being able to overspend, even if a prompt injection fools it.
An owner funds a NEAR vault contract and signs a mandate: which agent may ask, which accounts it may pay, how much per payment, per UTC day and in total, and until when. The agent holds a function-call key that can only call request_payment with zero deposit, so a stolen or fooled agent can ask but never spend. The contract checks every request itself (caller, revocation, policy version, expiry, allowlist, per-payment cap, replay, budgets), sends the transfer only when all pass, and records settlement in a private callback.
Proof on NEAR Testnet: 2 settled payments, one made by an AI agent through our MCP server, and 8 attacks refused by the contract (overspend, wrong recipient, a look-alike account, replay, old policy version, expired request, wrong signer, revoked mandate), each with zero transfers. Against a real model on NEAR AI Cloud, 19 prompt injections led to 0 escaped payments; every reply is signed inside NEAR AI's TEE and the attestation passed the official verifier. Judges can try their own injection live.
For builders: an MCP server that checks a recipient, catches look-alike accounts and dry-runs a payment before signing, plus an offchain guard.
Everything is verifiable: the site rechecks each transaction from public archival RPC, GitHub reruns the checks daily, and the deployed WASM rebuilds byte for byte from the source. Testnet only; the contract is not audited.